Efficient and Accurate Forecasting of Evolving Time Series from the Energy Domain

Datum

20.11.2014

Zeit

09:30 - 10:30

Sprecher

Dipl.-Medieninf. Lars Dannecker

Zugehörigkeit

Institut für Systemarchitektur, Professur Datenbanken

Sprache

en

Hauptthema

Informatik

Andere Themen

Informatik

Beschreibung

Continuous balancing of energy demand and supply is a fundamental prerequisite for the stability and efficiency of energy grids. This balancing task requires accurate forecasts of future electricity consumption and production at any point in time. Today's energy data management systems (EDMS) typically use statistical algorithms---called forecast models---that already provide accurate predictions. However, recent developments in the energy domain such as real-time intra-day trading and the integration of more renewable energy sources also require more efficient forecasting calculations and a rapid provisioning of forecasting results. Furthermore, todays EDMS fulfill a number of different tasks each exhibiting different requirements for the calculation of forecasts with respect to runtime and accuracy. Thus, it is necessary to flexibly adapt the forecasting process with respect to the needs of the current requests. In contrast, currently employed forecasting approaches are rather time-consuming and inflexible. One reason is the very expensive estimation of the forecast model parameters involving a large number of simulations in a search space that increases exponential with the number of parameters. We tackle these new requirements of improving the forecasting calculation efficiency and providing forecast process adaptability by introducing our novel online forecasting process. This process employs a special forecast model materialization in conjunction with a flexible and fast parameter estimation process to rapidly provide accurate results that are iteratively improved over time. EDMS may subscribe to the online forecasting process to retrieve improvements found during the process execution. In addition, they can adapt the progression of the forecasting calculation by defining runtime constraints and accuracy targets. With that we are able to equally server requests that require results in a limited time frame or that are targeting the best possible accuracy. The online forecasting process is complemented by further optimizations on the logical as well as on the physical layer. Our optimizations on the logical layer improve the efficiency of the parameter estimation independently of the data organization and the employed forecast model. As a first approach, we introduce our context-aware forecast model repository that materializes previously used forecast models and their parameters in conjunction with information about the time series context that was valid during the time the model was used. We may then provide appropriate starting points for future forecasting calculations by reusing models that produced accurate results in a context similar to the current one. Furthermore, for some use cases it is beneficial to consider context information directly within the forecast models. Especially when predicting renewable supply, information about the weather are very important. However, including context information typically means to add further parameters to the forecast model, which increases the efforts for the parameter estimation. To solve this issue, we introduce an integration framework that optimizes the handling of context information and reduces the additional efforts when considering them. Finally, we optimize the calculation of forecasts in hierarchical environments. Instead of simply aggregating time series, we propose a forecast model aggregation that eliminates the need for estimating the forecast model parameters on higher hierarchical levels. The physical optimizations introduced in this thesis directly provide an efficient way for forecast models to access time series values. For this purpose, we introduce an access-pattern-aware storage approach that exploits the memory access patterns of the used forecast models to physically layout the data for sequential access and high spatial locality. With that we substantially reduce the negative influence of memory latency and bandwidth, while at the same time improving the utilization of the different cache levels. In addition, we propose a special parallelization approach for multi-equation forecast models. Overall, with the help of our online forecasting process in conjunction with the optimizations on the logical and on the physical layer, we target to enable accurate forecasting of evolving time series in the face of the new requirements in the energy domain.

Letztmalig verändert: 20.11.2014, 08:55:02